kibitzer-s2-shaw-100m

kibitzer-s2-shaw-100m

Attention-first chess policy/value transformer (Kibitzer, S2 tier), trained on 100M positions from Lichess Elite games (2025-01..10) with shaw relative position encoding. Position-only (context_window=1); predicts AlphaZero-style moves + a bounded value head.

Architecture

• 15,223,105 parameters - attention-first (every trunk block is causal attention, attention_every=1)
• d_model=256, trunk_layers=10, n_heads=8, encoder_heads=8
• Position encoding: shaw (relative, Chessformer-style)
• Policy: 4,672 AlphaZero-style move logits; Value: side-to-move scalar in [-1, 1]

Training

• 100,000,000 positions, single pass, batch 128, AdamW, LR 1.5e-4 (constant), fp32
• Held-out eval (Lichess Elite 2025-11): top-1 49.45%, value MSE 0.650

Play strength (PUCT search, 256 sims, N=20 vs Stockfish UCI_Elo)

• vs SF-1900: 0.900 (17W/2D/1L); vs SF-2100: 0.700; vs SF-2300: 0.833
• Estimated Elo ~2280-2470 (single-anchor ~2282; 3-level gauntlet ~2470; wide CI, N=15-20/level)
• In-loop SF-1900 @ 64 sims scaling: 20M=0.25 -> 40M=0.35 -> 60M=0.65 -> 80M=0.775 -> 100M=0.825

Source

https://github.com/Mantissagithub/kibitzer/tree/clean-rebuild